CASE STUDY
PRISM
AI-ready research intelligence platform on a thousands-strong TBI cohort
THE SITUATION
Years of clinical evidence, locked behind a wall of cleanup
PRISM had spent years collecting longitudinal traumatic brain injury data on a cohort several thousand patients deep. The science was world-class. The infrastructure underneath it was a tangle of exports, spreadsheets, and legacy database snapshots no engineer wanted to touch.
Every meaningful question took a week of preparation before anyone could even ask it. The team needed a foundation they could actually do research on, not a backlog of cleanup tickets.
BEFORE
A week of cleanup before any real question could be asked.
AFTER
Researchers query the cohort in plain English. Engineering is no longer the bottleneck.
WHAT I BUILT
An AI-ready research workspace co-designed with the clinicians
I worked beside the researchers, not just for them, to model the cohort the way the clinical team actually reasons about it. Patient, visit, and outcome entities live in a typed schema tuned for both day-to-day querying and the heavy analytical cuts the lab needs for paper-grade work.
On top of that foundation, AI assists where it earns its place. Plain-English queries against the cohort, automated reconciliation that surfaces conflicts instead of hiding them, and a clean analytical surface ready for downstream LLM and ML work.
Clinical-first schema
Entities and relationships reviewed line by line with the lead researcher before a single record moved.
Reconciled ingestion
Fragmented historical sources flow through pipelines that flag discrepancies for review instead of silently picking a winner.
AI-ready surface
A typed analytical layer the clinical team queries in plain English, and that downstream ML models can train against without rework.
THE OUTCOMES
Used every week by the people it was built for
Thousands
patient cohort, fully unified
Weekly
plain-English queries by the clinical team, no engineer in the loop
Several
manuscripts in flight on insights this platform surfaced
STACK
- Postgres
- Snowflake
- Python
- OpenAI
- dbt
- Airflow
